Bayreuth GmbH:
Thermoset formulations are made by trial and error and by SME. Sometimes made by Design of experiments and in future by ML.
Bayesian optimization:
- Select random experiments
- Perform the experiments
- Train a model (Gaussian process) – it is not very accurate but can do very easy “virtual” experments
- Then create a Acquisition function (combine Prediction and Sigma)
- It is utility function and then feed it back to the model to be trained.
Eg: Bio based formulations – such as epoxy amine reactions by trying to determine the glass transition temperature.
Also – machine learning to develop flame retardant formulations.
Then it is possible to do multi-parameter optimization.
But beyond formulations, it is possible to do it in foam processing – Foam extrusion.
Foam extrusion – PLA bead foam process involving 20 time dependent monitored parameters. These can be optimized by determining the optimal Melt pressure…The training was done by iteration and then learning what had happened in the past.
Minimize air consumption in steam chest molding of bead foams. This is optimized to minimize the use of volume of air.
Minimize the batch foam density of styrene-MMA copolymers – this was optimized the processing characteristics for the autoclave processes. Sometimes this may lead to unexpected processes.
Composite process such as Force process for various molding operations such as tape laying where there is possibility of wastage. This was done by forming a digital twin.
It needs combination of machine learning, simulation and advanced testing since dynamic tests may end of costing significant resources in time and effort.
This can also be used for Fiber volume in composites by using CNN training and XGBoost Training.
Recycling: There is no single way for recycling of plastics. It requires collection – pretreatment washing and other steps required for chemical recycling. For plastic recycling washing can be optimized too.
SpecReK – mechanical recycling. You need to optimize for final product quality and that needs to optimized by multiple processes such as ML to maximize melt viscosity.
Bead foam molding : It is done with steam chest molding but RF technology can allow directly heating of foam beads. This requires optimization of bead manufacture can be done by optimizing the multi physics process with COMSOL. Process is predictable by physical relationships and can be modeled by simulation tools.
https://epub.uni-bayreuth.de/id/eprint/8759/1/1-s2.0-S2666827024000859-main.pdf
https://github.com/Polymer-Engineering-University-Bayreuth/FiberVolumeContent_ConvNet
Dr Sun – Changchun institute of Applied Chemistry
Majority of polymers have never been explored – e.g. Polyethylene, polypropylene are the most common ones. But with combinatorial methods possible the majority have not been discovered. Traditional polymer design does not scale. AI is needed to generate polymers
How does AI understand polymers using NLP tools such as structure, pattern and function.
Polymer language: Polymer can be considered as a graph. The chemicals can be described by physical -chemical descriptors such as molecular weight.
One way of doing this is to use PSMILES: representing polymers as sequences.
Polymers can thus be represented as ordered strings with models learns recurring structural motifs and valid structures follow underlying constraints. This enables tokenization to be leveraged at the SMILES level.
PolyNC – integrates natural and polymer language.
Python libraries are available but commercial packages may be pricey.
